Gemma 4 26B A4B It

Overview. Gemma 4 26B A4B IT is Google DeepMind's instruction-tuned, sparse Mixture-of-Experts model, released in April 2026 under an open Apache 2.0 licence. It totals 25.2B parameters but activates only 3.8B per token across 8 of 128+1 experts, pairing MoE efficiency with a 262,144-token context window and text, image, and video input. Developers who want an open-weights, multimodal model they can self-host for interactive-latency chat and moderate-complexity tasks would reach for this variant.

Key strengths. Domain scoring from Artificial Analysis puts this model at roughly a 39.3 coding index, meaningfully ahead of its 11 agentic index — real capability on coding tasks that doesn't yet extend to complex multi-step tool orchestration. It ships with native function-calling, a system role, structured-output support, and a configurable thinking mode via the model's <|think|> control token, so teams can opt into deeper reasoning only when a task calls for it. The sparse MoE design lets it deliver this feature set while touching just 3.8B active parameters per token, keeping inference lighter than a dense model of similar total size.

When to use it. Use this model for interactive chat, coding assistance, or multimodal (image and video) understanding where moderate complexity and quick turnaround matter more than frontier reasoning depth. Enable its optional thinking mode when a request needs extra deliberation, and lean on its native tool-calling and structured-output support for lightweight agentic workflows rather than long autonomous chains. It also suits teams that specifically want open weights they can deploy on their own infrastructure.

Limitations. Its coding and agentic index scores trail the family's larger 31B IT sibling, so it isn't the strongest choice for heavy code generation or complex multi-step agentic tasks. Reported max output length has varied across sources (most recently 16,384 tokens) even though the context window itself is a fixed 262,144 tokens, so single-turn generations shouldn't be assumed to run that long. Training data has a January 2025 cutoff.

Sovereignty note. Google is a United States-based creator with confirmed CLOUD Act exposure, a factor to weigh even though this model's weights are openly licensed under Apache 2.0. Because the weights are open, Gemma 4 26B A4B IT can be deployed on Sovereign infrastructure of your choosing, giving teams control over data residency independent of Google's own hosting.